Google’s September 2026 release of Gemini 3.8 Live Extended Thinking marks a departure from the industry standard of opaque, black-box inference. By enabling the model to reason and speak simultaneously, Google has shifted the transparency needle from static, post-hoc disclosures to a dynamic, real-time visibility model. This capability allows the system to narrate its thought process as it unfolds, fundamentally altering how users interact with and evaluate the reliability of agentic AI systems.
At its core, Extended Thinking is designed to handle complex workflows without breaking conversational flow. The model utilizes early verbal cues — such as the phrase, “Let me check that” — to acknowledge user prompts while it simultaneously executes background tasks. This is supported by asynchronous function calling, which allows the model to perform API and tool interactions while continuing to stream audio responses. With a 128K token context window and 64K token output capacity, the architecture is built to maintain coherence during multi-step reasoning chains.
The benchmarks are concrete. Extended Thinking ranks first on the Artificial Analysis Speech-to-Speech Quality Index at 82.6, scores 97.7% on Big Bench Audio, and leads agentic task completion at 68.6% on τ-Voice. Pricing sits at $0.005 per minute for audio input and $0.018 per minute for audio output — competitive enough for developers to build production voice agents at scale. Google has made the model available through its Gemini Live API, Google AI Studio, and across consumer products including Gmail, Docs, and Keep.
The shift toward visible inference serves as a new mechanism for error detection and user accountability. By exposing the reasoning chain through live audio, users are no longer passive recipients of a final output; they become active monitors of the model’s logic. If a model begins to drift or misinterpret a request, the verbal narration provides an immediate opportunity for the user to intervene. This creates a more intuitive accountability loop, moving away from the “take it or leave it” nature of traditional LLM interactions and toward a collaborative, verifiable process.
The industry is currently navigating a structural tension between process-level and interaction-level visibility. While Chinese initiatives, such as the Xiaomi MiMo-V2.6 live dashboard, have focused on providing visibility into the training process — real-time cost telemetry, token throughput, benchmark performance during an active RL run — Google’s approach emphasizes interaction-level transparency. The reasoning chain is visible to the end user as it happens, not after the fact. This divergence suggests that the industry has yet to settle on a unified standard for what “transparency” actually entails. One camp prioritizes the provenance of the model’s knowledge; the other prioritizes the integrity of the model’s real-time decision-making.
The practical implications for trust are significant. By narrating its progress, the model provides a window into its internal state, which may help mitigate the “black box” problem that has long hindered enterprise adoption. When a user can hear the model working through a problem, they gain a better understanding of the model’s limitations and its path to a conclusion. This aligns with broader accountability architecture trends, such as the OpenAI misalignment disclosure framework, which seeks to formalize how models communicate their own potential for failure.
Despite the technical sophistication, the model is not a panacea for AI reliability. The reliance on verbal narration assumes that users have the time and expertise to monitor these reasoning chains effectively. Furthermore, the integration of SynthID watermarking — an imperceptible watermark woven directly into the audio output — raises questions about how verifiable transparency will coexist with the need for secure, authenticated AI interactions. All audio generated by the model carries this watermark, a measure designed to prevent misinformation but one that adds another layer of complexity to the trust equation.
The trajectory of Extended Thinking will be defined by its adoption among enterprise partners like Salesforce, Genspark, and Lumeris, and by the developer ecosystem forming around the Gemini Live API. Partners including Agora, Fishjam, LiveKit, LangChain, Pipecat, Vercel, and Vision Agents are already building custom voice agents on the platform. The open question is whether real-time reasoning narration becomes a durable feature of trustworthy AI systems or a transitional measure until more robust, autonomous reasoning eliminates the need for human oversight of the inference chain.